What Are the Issues Surrounding Medical Transcription?

A service quoting a 24-hour turnaround is describing its target, not its capacity. The delay a practice actually experiences arrives when several clinicians dictate on the same afternoon, or when a holiday compresses a week of work into three days. The transcript that matters - the one needed before a referral goes out - is the one caught in that queue.

The useful question to a prospective provider is not "how fast are you" but "what happens when you are busy". A provider that cannot answer it is quoting a best case. A provider that offers tiered turnaround at stated prices is telling you what it can commit to, which is a different and more reliable kind of answer.

The headline figure - 99%, 99.9% - is an average across a corpus somebody else chose. It tells you nothing about the field you are worried about. A 99% accurate transcript of a 1,500-word consultation contains roughly fifteen errors, and whether that matters depends entirely on whether they fall in the social history or in the medication list.

The errors that cause harm share a characteristic: they are fluent. A phonetically similar drug name substituted for the right one reads correctly. A dose normalised toward a more usual value looks plausible. A dropped negation - "no history of" becoming "history of" - inverts the meaning while remaining grammatical. None of these looks wrong on the page, which is why a single accuracy percentage is the least useful thing a provider can tell you.

Audio containing protected health information disclosed to a transcription provider makes that provider a business associate under 45 CFR 164.308(b)(1), and the agreement has to exist before the disclosure rather than after it. This is the part most often handled informally, and it is not a formality: the practice remains accountable for what happens to that audio.

The questions worth asking are concrete. Where is the audio stored, and for how long? Who can access it? Is it encrypted at rest as well as in transit? Can it be deleted on request, and is deletion confirmed? A provider that answers these specifically is a provider that has thought about them.

A transcript costs a few dollars. Correcting a chart entry that reached a referral, a payer or a plaintiff costs considerably more, and the time spent is a clinician's rather than an administrator's. That asymmetry is why the cheapest option is often not the cheapest option.

It also explains why review should be a decision rather than a default in either direction. A transcript of a staff meeting does not need a second pair of eyes. A clinical dictation that will be signed into a chart probably does, and the price difference between the two is small against the cost of getting one wrong.

ScribeForms returns a confidence score and the transcript quote behind every extracted field, rather than one figure for the job. A reviewer's attention goes to the fields the model was least certain about, which is where the errors actually are. We publish per-field confidence precisely because a headline accuracy number is unverifiable and the per-field version is checkable.

Human review is a per-job choice with a stated rate difference, and the finished document records which mode produced it. So an AI-only transcript says so plainly rather than staying silent about whether anyone looked at it - and an auditor or a defence can tell the two apart.

A BAA is available on request, data is encrypted with AES-256 in transit and at rest, and access is logged. We are not SOC 2 certified; that audit is in progress and we will publish the report when it exists rather than before.